Tracking by an Optimal Sequence of Linear PredictorsKarel Zimmermann, Tomas Svoboda, Jiri Matas 
Center for Machine Perception Czech Technical University Prague http://cmp.felk.cvut.cz/ 
MoviesMotion blur (images) 
DownloadAbstractWe propose a learnable tracking method, where the computational complexity is explictly taken into account. The computational complexity, defined as the total number of pixels used for linear motion prediction, is minimized during an offline learning stage. Learning algorithm, based on linear and dynamic programming, searches for an optimal predictor, given a range of admissible motions, desired accuracy and a set of training examples. As a result, the globally optimal sequence of predictors covering the object is delivered. The local motion is determined only by a few hundreds of multiplications resulting in tracking with a fraction of the processing power of a desktop PC. The global motion of the object is robustly estimated by Ransac algorithm from some local motions. We also make explicit the tradeoff between the number of estimated local motions and the number of Ransac iterations. The optimal predictor is applied to realtime object tracking and its performance superiority and robustness are experimentally demonstrated. Support
References
